TSKD: A Teacher-Student Model for Session-Based Recommendation Using Knowledge Distillation

Session-based recommendation (SBR) aims to predict the next interaction of anonymous users based on their limited historical behaviour. However, most existing SBR methods still suffer from data sparsity issues due to short-term interactions. To alleviate the above issues, we propose a teacher and student distillation framework called TSKD for SBR. TSKD characterizes user intent by building multiple graphs. Specifically, a student graph (cross-session graph) and a teacher graph (item attribute similarity graph) are first constructed to supplement the session data. Then, graph neural networks (GNNs) and knowledge distillation (KG) are combined to leverage multiple graph structures for self-ensemble graph learning, effectively utilizing and learning item relationships from the student and teacher model. Furthermore, a multiple graph consistency constraint is introduced to align the student and teacher graphs in the feature embedding space, which enables the student model to learn more effectively from the teacher model by incorporating additional item relationships, leading to improved performance. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of TSKD.